Controller for a power device, power assembly comprising such a controller, and method for operating a power device

Abstract

The invention relates to a controller (3) for a power device (5), said controller (3) being designed to determine an actuation manipulated variable vector comprising manipulated variables for actuating the power device (5), comprising - a first optimization module (7) which is designed to determine a first optimized manipulated variable vector using a first non-evolutionary algorithm, - a second optimization module (9) which is designed to receive the first optimized manipulated variable vector from the first optimization module (7) and determine a second optimized manipulated variable vector by means of a second evolutionary algorithm using the first optimized manipulated variable vector, and - a control module (11) which is designed to determine the actuation manipulated variable vector on the basis of the second optimized manipulated variable vector.

Description

[0001]Rolls-Royce Solutions GmbH DESCRIPTION Control device for a power device, power arrangement with such a control device and method for operating a power device The invention relates to a control device for a power device, a power arrangement with a power device and such a control device and a method for operating a power device. In order to find optimized manipulated variables for controlling a power device, particularly with regard to achieving certain operating goals, control devices are typically designed to implement classic, deterministic methods or methods derived therefrom. By means of such methods, particularly when they rely on gradients, only a local optimum within a characteristic map can be found with relative certainty. This can lead tothat a power device is operated at one and the same operating point at different times with different manipulated variables. While this is not necessarily disadvantageous with regard to the operating objectives to be achieved, it is problematic in sensitive areas where the operation of the power device must be certified. As a rule, in such areas, for safety reasons, it is already required that the power device exhibits a clear operating behavior, in particular a clear assignment of manipulated variables to operating points in the characteristic map. While it is fundamentally possible to search for a global optimum in the characteristic map using simple extensions such as branch-and-bound, starting from several starting values, such methods involve a high numerical effort, which is hardly feasible in real time on a typical control device, especially not during the runtime of the power device.can be provided. The invention is therefore based on the object of creating a power device, a power arrangement with a power device and such a control device and a method for operating a power device, wherein the aforementioned disadvantages are at least reduced, preferably not occurring. The object is achieved by providing the present technical teaching, in particular the teaching of the independent claims and the preferred embodiments disclosed in the dependent claims and the description. The object is achieved in particular by creating a control device for a power device, wherein the control device is configured to determine a control manipulated variable vector which has - as components - manipulated variables for controlling the power device. The control device has a first optimization module which is configured to, by means of a first,non-evolutionary algorithm to determine a first optimized manipulated variable vector. The control device also has a second optimization module configured to receive the first optimized manipulated variable vector from the first optimization module and to determine a second optimized manipulated variable vector by means of a second evolutionary algorithm using the first optimized manipulated variable vector. Finally, the control device has a control module configured to determine the control manipulated variable vector as a function of the second optimized manipulated variable vector. By combining two different optimization methods, a first of which uses a non-evolutionary algorithm, and a second of which uses an evolutionary algorithm, various advantages can be combined with limited numerical effort: The non-evolutionary algorithm enables very precise calculation,in particular by searching for a local optimum, while the evolutionary algorithm extends the search in a simple and numerically manageable manner to areas of the characteristic map away from the local optimum, thus opening up the possibility of finding a global optimum. The combination of the two algorithms thus makes it possible, in particular, to search for a global optimum in real time, i.e., during runtime during the operation of the power device, particularly on a conventional control device, i.e., one that has not been specifically upgraded in terms of its computing power. This allows, not least, the operation of the power device to be cost-effective – not only with regard to achieving operating objectives such as the lowest possible fuel consumption or compliance with statutory emission targets,but also with a view to certification. In particular, the second optimized manipulated variable vector is determined in real time, in particular during the runtime of the control device and in particular during the runtime of the power device controlled by the control device. In particular, the control manipulated variable vector is determined in real time, in particular during the runtime of the control device and in particular during the runtime of the power device controlled by the control device. In the context of the present technical teaching, a power device is understood in particular to mean a device that is configured to provide power, in particular electrical and / or mechanical power,or to convert or consume power. The power device can thus be designed in particular as a power supply device or as a power conversion device. A power supply device is understood in particular to be a device that provides power, in particular electrical and / or mechanical power, using electrical, mechanical, chemical or electrochemical energy - or another form of energy. A power conversion device is understood in particular to be a device that uses or consumes power, in particular electrical or mechanical power, in particular to convert or store energy, for example to use electrical energy to produce chemical energy in the form of certain substances such as hydrogen or methanol,or electrochemical energy. In particular, the power device can be an internal combustion engine, an internal combustion engine-generator combination device, i.e., a genset, a fuel cell, an energy storage device, in particular a battery, or an electrolyzer. However, the power device can also be a larger, more complex system, for example, comprising a plurality of the aforementioned devices, or in particular also a data center or a microgrid. In particular, the power device can also be a controllable or regulatable load on an electrical network. The control device is in particular configured to operate the power device. In one embodiment, the control device is configured to operate an internal combustion engine, an internal combustion engine-generator combination device, i.e., a genset, a fuel cell, an energy storage device,in particular a battery, an electrolyzer, a data center or microgrid, or another controllable or regulatable load on an electrical network. In particular, the control device is configured to operate the power device with the control manipulated variable vector, that is to say in particular to control it. A vector, specifically a manipulated variable vector, in the context of the present technical teaching is understood to mean in particular a combination of at least one manipulated variable, regardless of how the at least one manipulated variable is depicted or represented within the control device. In particular, any desired number or combination of manipulated variables can be understood or depicted as a vector with at least one component, wherein the number of vector components corresponds to the number of manipulated variables. In the context of the present technical teaching, a manipulated variable is understood to mean a variablewhich is suitable for controlling an actuator of the power device. It is possible for the manipulated variable to be directly suitable for controlling the actuator, for example, because it is given as a specific voltage or current that can be directly applied to the actuator in order to control it; alternatively, it is possible for the manipulated variable to be suitable for deriving, in particular calculating, at least one further variable for directly controlling the actuator. For example, the manipulated variable can be a fuel mass to be introduced into a combustion chamber of an internal combustion engine, which can be converted into at least one control variable for controlling an injector. The at least one actuator can be an actuator or actuator of the power device. In particular, the actuator can be an actuator of an engine block of an internal combustion engine, for example, an injector.a valve or a flap. However, the actuator can also be an actuator outside the power device, in particular outside an engine block, for example, an actuator intended to influence an externally provided cooling circuit, for example, a valve, a pump, or the like, or an actuator of a transmission or an electrical device to which the power device is operatively connected. In the context of the present technical teaching, a module is generally understood to mean, in particular, a mentally or physically delimitable or delimited functional unit that is configured to perform at least one specific function. This can be a separate computing device, a part of a computing device, a hardware structure, or a software structure.which is configured and provided to fulfill at least one specific function. In one embodiment, the control device is configured to detect, determine, or receive—in particular current—operating parameters of the power device, such as measured values, for example an ambient temperature or an ambient air pressure, target values ​​for operation, limit values, environmental boundary conditions, and optionally further parameters. In particular, the first optimization module is configured to use the operating parameters to determine the first optimized manipulated variable vector; in particular, the first optimized manipulated variable vector is determined as a function of the operating parameters. In one embodiment, the first optimization module is configured to receive the second optimized manipulated variable vector from the second optimization module, and the second optimization module is configuredto transmit the second optimized manipulated variable vector to the first optimization module. The first optimization module is configured, in particular, to determine the first optimized manipulated variable vector, in particular—within the scope of the iteration described below—a new first optimized manipulated variable vector, based on the received second optimized manipulated variable vector. According to a further development of the invention, the first optimization module is configured to determine the first optimized manipulated variable vector using a gradient method as the first, non-evolutionary algorithm. In particular, in this way, a local optimum in the characteristic map can be found with very high accuracy by the first optimization module. According to a further development of the invention, the second optimization module is configured to generate a predetermined number of test manipulated variable vectors,to use the first optimized manipulated variable vector as a further test manipulated variable vector, so that a first set of manipulated variable vectors with the cardinality of a predetermined total number is formed. The first set of manipulated variable vectors comprises the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector. The second optimization module is further configured to determine the second optimized manipulated variable vector from the first set of manipulated variable vectors using the second evolutionary algorithm. Thus, the second optimization module can use the evolutionary algorithm to easily and computationally inexpensively also achieve manipulated variable vectors outside of the local optimum possibly found with the first optimized manipulated variable vector, thus contributing toto find, if possible, a global optimum and, in particular, a unique control variable vector for a specific operating point—especially with regard to certification. In particular, the test control variable vectors are randomly generated in one embodiment, in particular from uniformly distributed random numbers, with a random number being generated for each component of a test control variable vector. In one embodiment, the test control variable vectors are generated for this purpose from uniformly distributed random numbers, which can, in particular, assume values ​​from 0 to 1, i.e., from the closed interval [0, 1]. The first optimized control variable vector in this embodiment is also standardized component-wise such that its components have values ​​from 0 to 1. In particular, the value range of each control variable is limited to the interval [0,1]. The control manipulated variable vector obtained later also has components with values ​​from the interval [0,1]. It is then advantageously denormalized component-wise for controlling the assigned actuators, whereby the interval [0,1] is again mapped to the actual value range of the assigned manipulated variable. In particular, the first set of manipulated variable vectors contains the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector. In particular, the first set of manipulated variable vectors consists of the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector. According to a further development of the invention, it is provided that the second optimization module is configured to - as the second,evolutionary algorithm – a) to form the first set of manipulated variable vectors from the predetermined number of randomly generated test manipulated variable vectors and the first optimized manipulated variable vector received from the first optimization module as a further test manipulated variable vector; b) to evaluate each test manipulated variable vector of the first set using a cost function; c) to form a second set of descendant manipulated variable vectors with the cardinality of the predetermined total number from a predetermined parent number of the test manipulated variable vectors of the first set that were best evaluated using the cost function; d) to change the descendant manipulated variable vectors with the exception of a predetermined minority of elite manipulated variable vectors – in particular using a uniformly distributed noise process – whereby a third set of manipulated variable vectors with the cardinality of the predetermined total number is formed,where the third set comprises the modified offspring manipulated variable vectors and the elite manipulated variable vectors, and e) to determine the test manipulated variable vector best evaluated using the cost function as the second optimized manipulated variable vector. In particular, using this procedure, it becomes possible to search in a structured manner beyond the first optimized manipulated variable vector for manipulated variable vectors that may be better with regard to the cost function, and to determine a manipulated variable vector as the second optimized manipulated variable vector that is better in terms of the cost function, i.e., more cost-effective than the first optimized manipulated variable vector. Of course, it is also possible,that the first optimized manipulated variable vector is already best evaluated with regard to the cost function and is output as the second optimized manipulated variable vector. In particular, the test manipulated variable vectors are ranked in step b) based on the evaluation using the cost function, with the – best – test manipulated variable vector, to which the lowest costs are assigned according to the cost function, having the highest rank, and the test manipulated variable vector, to which the highest costs are assigned according to the cost function, having the lowest rank. The costs assigned to the respective test manipulated variable vector, and thus the assigned rank, are subsequently retained in the further steps and also assigned accordingly to the descendant manipulated variable vector assigned to the test manipulated variable vector, if applicable also to the assigned elite manipulated variable vector, as well as to the assigned modified descendant manipulated variable vector.thus inherited. In one embodiment, in a first sub-step b1) of step b), the test manipulated variable vectors are projected onto a permissible normalized range, where for each manipulated variable, i.e. for each component x, j of the test manipulated variable vectors x each have a permissible range 0 <= LBU j <= x j <= UBU j <= 1 (1) within a lower limit LBU j and an upper limit UBU j defined interval [LBU j ,UBU j ] for the respective component x j The lower limit LBU j and the upper limit UBU jcan be redetermined, in particular, for each iteration step of the iteration explained below, in particular depending on the – in particular current – ​​operating parameters of the power device. In this way, it can advantageously be taken into account, in particular, that not every arbitrary value of a manipulated variable can be assumed at any time or – in particular to protect the power device – may be assumed. Alternatively or additionally, in a second sub-step b2) of step b), the operating parameters of the power device are received by the second optimization module. These operating parameters are used, in particular, in a third sub-step b3) of step b) to calculate the cost function.In one embodiment, in step b), particularly in the third sub-step b3), the test manipulated variable vectors are evaluated using the following procedure: For each test manipulated variable vector, it is checked whether it satisfies at least one predetermined restriction—for example, with regard to a maximum permissible peak pressure in the combustion chamber or a maximum permissible emission value. If the tested test manipulated variable vector satisfies the at least one predetermined restriction, it is evaluated by calculating the cost function; if, however, the tested test manipulated variable vector violates at least one restriction selected from the at least one predetermined restriction, a maximum cost value of the cost function is assigned to this test manipulated variable vector.This advantageously results in those test manipulated variable vectors that violate constraints receiving the worst possible rating and being ranked lowest in the test manipulated variable vectors. The maximum cost value can be a predetermined value or it can be calculated using a predetermined algorithm, in particular by adding a violation contribution to the cost function. It is possible for the at least one predetermined constraint to be changed within the iterations described below, in particular if operating parameters of the power device change. In particular, the at least one constraint can be redetermined for each iteration step of the iteration explained below, in particular depending on the operating parameters.In one embodiment, the ranking of the test manipulated variable vectors is determined in step b), in particular in a fourth sub-step b4), using a sorting algorithm, in particular merge sort, depending on the evaluation using the cost function. In one embodiment, the predetermined parent number is smaller than the predetermined total number. In one embodiment, the predetermined total number is 10; alternatively or additionally, the predetermined parent number is 4.In one embodiment, in step c), the second set of offspring manipulated variable vectors is formed from the predetermined parent number of the test manipulated variable vectors of the first set that are best evaluated using the cost function in the following way: First, in a first c-sub-step c1), first offspring manipulated variable vectors are generated from the test manipulated variable vectors that correspond to the predetermined parent number and are best evaluated according to the ranking in the following way: A offspring manipulated variable vector with the rank in the ranking corresponding to the predetermined parent number is generated by calculating, in particular averaging, the test manipulated variable vector with the rank corresponding to the predetermined parent number with the next higher test manipulated variable vector in the ranking, i.e. the offspring manipulated variable vector has component-wise mean values ​​of the corresponding components of the two test manipulated variable vectors that are calculated with each other.Another offspring manipulated variable vector with a rank corresponding to the predetermined parent number reduced by one is generated by similarly averaging this test manipulated variable vector with the next higher-ranked test manipulated variable vector. This process continues until the highest-ranking offspring manipulated variable vector is finally generated by equating it to the highest-ranking test manipulated variable vector. In particular, this is done according to the following equations: ^^ ^^ ^^ ^ ^,1 = ^^ ^^,1 (3) for all components j, where ^^ ∈ { 1, … , ^^ } describes the rank in the ranking with the predetermined total number G, E is the predetermined parent number, ^^ ^^ ^ ^, ^^ the component j of the offspring manipulated variable vector x N with the rank denoted by i, and ^^ ^^ ^ ^, ^^is the component j of the test manipulated variable vector xT with the rank denoted by i. The rank of a manipulated variable vector within the ranking is higher, the lower the value of i. In particular, the highest-ranking manipulated variable vector is assigned the value i = 1, whereas the lowest-ranking manipulated variable vector is assigned the value i = G. For example, if the predetermined number of parents is 4, the fourth-ranked offspring manipulated variable vector is generated by averaging the fourth-ranked test manipulated variable vector with the third-ranked test manipulated variable vector; the third-ranked offspring manipulated variable vector is generated by averaging the third-ranked test manipulated variable vector with the second-ranked test manipulated variable vector;the second offspring manipulated variable vector is generated by averaging the second-ranked test manipulated variable vector with the first-ranked – i.e. best or highest-ranking – test manipulated variable vector;Finally, the first offspring manipulated variable vector in the ranking is generated by setting it identically to the first – best – test manipulated variable vector in the ranking. In particular, the highest-ranking – best – test manipulated variable vector is used as the highest-ranking offspring manipulated variable vector. Then, in a second sub-step c2), second offspring manipulated variable vectors are generated such that, starting from the rank corresponding to the predetermined parent number increased by one, the offspring manipulated variable vectors are each selected identically to the previously generated first offspring manipulated variable vectors with the rank each decreased by the predetermined parent number, until the predetermined total number is reached, in particular according to the following equation: If, for example, the predetermined total number is 10 and the predetermined parent number is 4, the fifth offspring manipulated variable vector in the ranking is now selected to be identical to the first offspring manipulated variable vector, the sixth offspring manipulated variable vector is selected to be identical to the second offspring manipulated variable vector, the seventh offspring manipulated variable vector is selected to be identical to the third offspring manipulated variable vector, the eighth offspring manipulated variable vector is selected to be identical to the fourth offspring manipulated variable vector, the ninth offspring manipulated variable vector is selected to be identical to the fifth offspring manipulated variable vector - and thus also identical to the first offspring manipulated variable vector - and the tenth offspring manipulated variable vector is selected to be identical to the sixth offspring manipulated variable vector - and thus also identical to the second offspring manipulated variable vector.In particular, the second set of descendant manipulated variable vectors contains the predetermined total number of descendant manipulated variable vectors. In particular, the second set of descendant manipulated variable vectors consists of the predetermined total number of descendant manipulated variable vectors. In particular, the descendant manipulated variable vectors are mutated in step d). In particular, the highest-ranking descendant manipulated variable vectors corresponding to the predetermined subset U in the ranking are used as the elite manipulated variable vectors. In particular, the predetermined subset U is smaller than the predetermined total number G. In one embodiment, the predetermined subset U = 2; in this case, the elite manipulated variable vectors are the first, highest-ranking descendant manipulated variable vector and the second descendant manipulated variable vector directly following in the ranking.In one embodiment, in step d), the descendant manipulated variable vectors, with the exception of the predetermined subset U of elite manipulated variable vectors, are modified, in particular mutated, according to the following procedure, wherein the third set of manipulated variable vectors is formed: A modified descendant manipulated variable vector is formed from the corresponding descendant manipulated variable vector in the ranking by calculating its components with a random number, in particular a uniformly distributed random number, which is in particular selected from the interval [0,1], taking into account a weight factor, which is in particular selected from the interval [0,1]. In particular, the respective component is multiplied by the difference of the weight factor to 1, and the product of the random number and the weight factor is added to the result.Advantageously, the offspring manipulated variable vectors, with the exception of the predetermined minority of elite manipulated variable vectors, are modified using a uniformly distributed noise process. In one embodiment, the weighting factor is selected depending on the rank of the offspring manipulated variable vector to be modified within the ranking, in particular, the lower the rank of the offspring manipulated variable vector, the higher the weighting factor. Higher-ranking offspring manipulated variable vectors are thus modified less than lower-ranking offspring manipulated variable vectors, or—conversely—lower-ranking offspring manipulated variable vectors are modified more than higher-ranking offspring manipulated variable vectors. In one embodiment, the relationship between the weighting factor and position i in the ranking is linear; as explained above, higher values ​​of i correspond to a lower rank.Alternatively or additionally, in one embodiment, to determine the weighting factor, a distance - in particular a Euclidean distance - of the descendant manipulated variable vector to be changed from the first optimized manipulated variable vector is calculated, wherein the weighting factor is selected to be maximum, in particular equal to 1, if the distance of the descendant manipulated variable vector to be changed from the first optimized manipulated variable vector is smaller than a predetermined limit distance. The predetermined limit distance is 0.1 in one embodiment. If, on the other hand, the distance is greater than or equal to the predetermined limit distance, the weighting factor is selected - as described above - depending on the position of the descendant manipulated variable vector within the ranking. In particular, each component j of the descendant manipulated variable vectors xN to be changed is changed according to the following equation: ^^. ^^ = 1 − ^^ ^^ ^ ^, ^^ ( ^^ ^^ ) ^^ ^^, ^^ + ^^^^ ∙ ^^ ^^ ∀ ^^ ∈ { ^^ + 1, … , ^^ } , (5) where xm is the changed offspring manipulated variable vector, ai is the weight factor assigned to the respective offspring manipulated variable vector x to be changed au ^^ N s the interval [0,1] and ^^ ^^ is the newly generated, in particular uniformly distributed, random number from the interval [0,1] for each component of each descendant manipulated variable vector xN. The larger i is, the larger the value of ai. In one embodiment, the function a i(i) a linear function. It is possible that ai = 1 is additionally selected if the distance—in particular Euclidean—between the descendant manipulated variable vector to be changed and the first optimized manipulated variable vector is smaller than a predetermined limit distance, in particular smaller than 0.1. In one embodiment, however, a descendant manipulated variable vector is not changed if it is identical to the first optimized manipulated variable vector. Alternatively or additionally, such a descendant manipulated variable vector—in deviation from the procedure according to the distance—is changed depending on its rank, in particular according to equation (5), if it is identical to the first optimized manipulated variable vector.In a further embodiment, it is provided that a descendant manipulated variable vector identical to the first optimized manipulated variable vector and having a relatively high rank, for example - in particular parameterizable - from 1 to 4, is not changed, while a descendant manipulated variable vector identical to the first optimized manipulated variable vector and having a relatively low rank, for example from 5 to 10, is changed depending on its rank, in particular according to equation (5). In particular, this procedure can advantageously have an accelerating effect on the search for a global optimum, in particular in the vicinity of a local optimum. In particular, the third set contains the changed descendant manipulated variable vectors and the elite manipulated variable vectors. In particular, the third set consists of the changed descendant manipulated variable vectors and the elite manipulated variable vectors.In particular, in step e), the second optimized manipulated variable vector is output to the control module, which determines the control manipulated variable vector based thereon. In particular, in step e), the second optimized manipulated variable vector is additionally transferred to the first optimization module, which then determines a new first optimized manipulated variable vector based on the second optimized manipulated variable vector.According to a further development of the invention, the second optimization module is further configured to f) receive the new, first optimized manipulated variable vector from the first optimization module and to form a fourth set of manipulated variable vectors with the size of the predetermined total number by replacing a manipulated variable vector from the third set with the new, first optimized manipulated variable vector, and g) to repeat steps b) to f) with the fourth set obtained in step f) as the first set, in particular until a predetermined termination condition is met. In this way, the search for an optimal control manipulated variable vector is advantageously carried out iteratively, with new developments, in particular changed operating conditions, being taken into account by adding the respective new, first optimized manipulated variable vector.In one embodiment, the first optimized manipulated variable vectors are determined by the first optimization module using a first time cycle having a first sampling time, wherein the second optimized manipulated variable vectors are determined by the second optimization module, in particular the implementation of steps b) to f), using a second time cycle having a second sampling time, wherein the first sampling time is shorter than the second sampling time. In particular, the first sampling time in one embodiment is 10 ms; alternatively or additionally, the second sampling time is 100 ms. Thus, in particular, a first optimized manipulated variable vector is determined more often than a second optimized manipulated variable vector.Thus, not every newly determined first optimized manipulated variable vector is considered as a new first optimized manipulated variable vector in step f), but rather, for example, every tenth newly determined first optimized manipulated variable vector for the specific sampling times mentioned. In particular, as described above, the new first optimized manipulated variable vector is determined by the first optimization module based on the second optimized manipulated variable vector determined in the previous iteration run. This ensures that the more accurate first, non-evolutionary algorithm searches for the first optimized manipulated variable vector in the region of the characteristic map by identifying the position of a possible global optimum using the second, evolutionary algorithm. In this way, the advantages of both algorithms are cleverly combined, and as a result, a very accurate solution for the global optimum can be iterated out.In one embodiment, to form the fourth set, the lowest-ranking, worst-rated manipulated variable vector, in particular the manipulated variable vector assigned the rank i = G, is replaced by the new, first optimized manipulated variable vector. In particular, the fourth set contains the predetermined total number of manipulated variable vectors. In particular, the fourth set consists of the predetermined total number of manipulated variable vectors. In one embodiment, the predetermined termination condition is checked in step e), with step f) and thus also the subsequent steps g) and b) to e) only being performed again if the termination condition is not met. In one embodiment, the reaching of a predetermined maximum number of iterations is used as the predetermined termination condition. Alternatively or additionally, the reaching of a predetermined maximum iteration time is used as the predetermined termination condition.Alternatively or additionally, switching off the control device or the power device is used as the predetermined termination condition. In one embodiment, the maximum number of iterations can be 1 - in particular to ensure the real-time capability of the method - so that only exactly one run is performed. In other cases, in particular when more computing power is available or no real-time system is used, the maximum number of iterations can also be greater than 1. A maximum iteration time can be used in embodiments in which, in the event of an error, invalid species are generated to such an extent that the model calculates invalid values, which could otherwise lead to a massive increase in the runtime. A maximum iteration time can also be used if no real-time system is used, whereby the available computing capacity can vary.By specifying the maximum iteration time, it can then be ensured that no outdated solution is further calculated. According to a further development of the invention, the first optimization module is configured to determine the first optimized manipulated variable vector for a predetermined prediction interval. In particular, the first optimization module is configured to determine the first optimized manipulated variable vector for the end of the predetermined prediction interval. Thus, the first optimized manipulated variable vector is determined, as it were, for the future, in particular as a goal to be achieved. According to a further development of the invention, the control module is configured to determine the control manipulated variable vector by means of a model-based, predictive method, penalizing any deviation of the control manipulated variable vector from the second optimized manipulated variable vector.This advantageously ensures that the control variable vector determined by the control module develops in the direction of the second optimized control variable vector. In the context of the present technical teaching, a model-based predictive method is understood in particular to mean model predictive control (MPC). In particular, the control module is configured to calculate the control variable vector using a Gaussian process model. Alternatively or additionally, the first optimization module is configured to calculate the first optimized control variable vector using a Gaussian process model, in particular using the same Gaussian process model that the control module also uses.Gaussian process models are particularly suitable for modeling and / or controlling a power device: Compared to polynomial-based models, they are easier to adapt to new or changed data points in the application field, and they exhibit more suitable and physically correct behavior at the edge of the given parameter space. Compared to physical models, they require significantly less computational effort. Furthermore, they allow the direct use of test bench data, data from – especially high-precision – simulations or models, or even field data from the application field. Such a Gaussian process model is given in particular by stored data points (Xb,Yb), for example, obtained in test bench tests or from simulations or models, where Xb ^. x m in particular n input variables for m different operating conditions and with Yb ^ ℝ m x kIn particular, k output variables are specified for the m different operating states. The output variables Y b are in particular a subset of the at least one manipulated variable. Furthermore, the Gaussian process model is characterized by a predetermined calculation scheme for an expected value E x k and a variance Var(X u ) for input variables not included in the original data set for l different operating states X u ^ ℝ n x l given: ^^ ( ^^ ^^ ) = ^^ ( ^^ ^^ ) + ^^ ( ^^ ^^ , ^^ ^^ )( ^^ ( ^^ ^^ , ^^ ^^ ) + ^^ 2 ^^ )−1 ( ^^ ^^ − ^^( ^^ ^^ )) , (6) ^^ ^^ ^^( ^^ ^^ ) = ^^( ^^ ^^ , ^^ ^^ ) + ^^ 2 − ^^( ^^ ^^ , ^^ ^^ )( ^^( ^^ ^^ , ^^ ^^ ) + ^^ 2 ^^) −1 ^^( ^^ ^^, ^^ ^^ ) , (7) with a mean function m(X u ), a predetermined variance ^ ^ 2 , the identity matrix I, and a covariance function K, which depends on the Euclidean distance r between two points x1, x2 in the following way: with a predetermined distance parameter l and a predetermined signal variance ^^ ^^ . Thus, in equations (6) and (7) K(Xu,Xb) ^ ℝ l x m , K(Xb,Xb) ^ x mand The mean value function m(x) is preferably itself obtained as a Gaussian process model. In particular, a first Gaussian process model, also referred to as the basic grid, is first adapted to second existing data under at least one constraint derived from first existing data, i.e., test bench data or from data obtained from – in particular, highly accurate – models or simulations. In particular, input variables Xb are selected, and the associated output variables Yb are calculated such that a deviation of the expected value E(X) of the first Gaussian process model, which is determined by the input variables Xb and the output variables Yb,to the second inventory data is minimized while observing the constraint. Furthermore, for the purpose of determining the first Gaussian process model, it is preferably assumed that m(x) = 0 is the mean value function of the first Gaussian process model. The first inventory data encompasses a larger parameter space than the second inventory data. For example, for an internal combustion engine, it is possible for the first inventory data to be measured on a single-cylinder test bench, while the second inventory data are measured on the full engine or also on the single-cylinder test bench and, in the latter case, converted to the full engine – preferably using a simulation model. The constraint is preferably obtained as a trend, whereby, for example, it is determined whether certain parameters behave linearly or monotonically with respect to one another. If no such trend is determined, the constraint can be omitted.wherein the adaptation of the Gaussian process model to the second set of data is then also referred to as unlimited. The expected value of the first Gaussian process model thus obtained is then used in a next step as the mean function m(x) in a second Gaussian process model, into which the second set of data are now input as known input variables Xb2 and output variables Yb2. According to a further development of the invention, the control module is configured to determine a manipulated variable vector trajectory for the predetermined prediction interval and to determine a current manipulated variable vector of the manipulated variable vector trajectory as the control manipulated variable vector. In particular, this advantageously ensuresthat the control variable vector determined by the control module develops in the direction of the second optimized manipulated variable vector and thus in particular in the direction of the global optimum. According to a further development of the invention, the first optimization module is configured to calculate the first optimized manipulated variable vector in a stationary manner. This allows for a particularly simple and, with regard to computational costs, cost-effective calculation of the first optimized manipulated variable vector, since the dynamics of the system are not considered here. According to a further development of the invention, the control module is configured to calculate the control variable vector, in particular the manipulated variable vector trajectory, taking into account the dynamics of the power device. In this way, the dynamics of the system are taken into account in the control module,wherein advantageously, consideration of the dynamics in the first optimization module can be dispensed with. The object is also achieved by creating a power arrangement which has a power device and a control device according to the invention or a control device according to one or more of the previously described embodiments. In connection with the power arrangement, in particular, those advantages which were already described previously in connection with the control device are realized. In particular, the control device is operatively connected to the power device in order to control the power device. In one embodiment, it is provided that the power device is designed as an internal combustion engine, an internal combustion engine-generator combination device, i.e., a genset, a fuel cell, an energy storage device, in particular a battery, an electrolyzer, a data center or a microgrid.or as another controllable or regulatable load on an electrical network. The object is also achieved by providing a method for operating a power device, wherein a first optimized manipulated variable vector is determined by means of a first, non-evolutionary algorithm, a second optimized manipulated variable vector is determined by means of a second, evolutionary algorithm using the first optimized manipulated variable vector, wherein a control manipulated variable vector is determined as a function of the second optimized manipulated variable vector, and wherein the power device is operated with the control manipulated variable vector. In connection with the method, in particular those advantages are realized which have already been described above in connection with the control device and the power arrangement. In one embodiment, it is provided that the power device is an internal combustion engine,an internal combustion engine-generator combination device, i.e., a genset, a fuel cell, an energy storage device, in particular a battery, an electrolyzer, a data center or microgrid, or another controllable or regulatable load is operated on an electrical network. In one embodiment, the first optimized manipulated variable vector is determined using a gradient method as the first, non-evolutionary algorithm. In one embodiment, a predetermined number of test manipulated variable vectors are generated, and the first optimized manipulated variable vector is used as a further test manipulated variable vector.such that a first set of manipulated variable vectors with the cardinality of a predetermined total number is formed. The first set of manipulated variable vectors comprises the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector. The second optimized manipulated variable vector is generated by means of the second,evolutionary algorithm from the first set of manipulated variable vectors. In one embodiment, a) the first set of manipulated variable vectors is formed from the predetermined number of randomly generated test manipulated variable vectors and the first optimized manipulated variable vector received from the first optimization module as a further test manipulated variable vector; b) each test manipulated variable vector of the first set is evaluated using a cost function; c) a second set of descendant manipulated variable vectors with the cardinality of the predetermined total number is formed from a predetermined parent number of the test manipulated variable vectors of the first set that were best evaluated using the cost function; d) the descendant manipulated variable vectors are changed with the exception of a predetermined minority of elite manipulated variable vectors, wherein a third set of manipulated variable vectors with the cardinality of the predetermined total number is formed,wherein the third set comprises the modified offspring manipulated variable vectors and the elite manipulated variable vectors, and e) the test manipulated variable vector best evaluated by the cost function is determined as the second optimized manipulated variable vector. In one embodiment, f) a new first optimized manipulated variable vector is received from the first optimization module, and a fourth set of manipulated variable vectors with the cardinality of the predetermined total number is formed by replacing a manipulated variable vector of the third set with the new first optimized manipulated variable vector, and g) steps b) to f) are repeated with the fourth set obtained in step f) as the first set,in particular until a predetermined termination condition is met. In one embodiment, the first optimized manipulated variable vector is determined for a predetermined prediction interval. In particular, the first optimized manipulated variable vector is determined for the end of the predetermined prediction interval. In one embodiment, the control manipulated variable vector is determined by means of a model-based, predictive method, penalizing a deviation of the control manipulated variable vector from the second optimized manipulated variable vector. In one embodiment, a manipulated variable vector trajectory is determined for the predetermined prediction interval, and a current manipulated variable vector of the manipulated variable vector trajectory is determined as the control manipulated variable vector. In particular, the first optimized manipulated variable vector is calculated in a stationary manner in one embodiment. In one embodiment, the control manipulated variable vector, in particular the manipulated variable vector trajectory,calculated taking into account the dynamics of the power device. In particular, the control device according to the invention or a control device according to one or more of the previously described embodiments is configured to carry out the method according to the invention or a method according to one or more of the previously described embodiments. The description of the control device on the one hand and the method on the other hand are to be understood as complementary. In particular, the method comprises at least one step or a combination of steps,which has been described explicitly or implicitly in connection with the control device. The invention is explained in more detail below with reference to the drawing. Figure 1 shows a schematic representation of an embodiment of a power arrangement with an embodiment of a control device; Figure 2 shows a schematic representation of an embodiment of a method for operating a power device in the form of a flow chart; and Figure 3 shows a schematic representation of a detail of the method according to Figure 2. Figure 1 shows a schematic representation of an embodiment of a power arrangement 1 with an embodiment of a control device 3. The power arrangement 1 also has a power device 5, which is operatively connected to the control device 3.so that the control device 3 can control the power device 5. In the exemplary embodiment shown here, the power device 5 is designed as an internal combustion engine 6. However, it can also be designed, in particular, as an internal combustion engine-generator combination device, i.e., a genset, fuel cell, energy storage device, in particular a battery, electrolyzer, data center, or microgrid, or as another controllable or regulatable load on an electrical network. The control device 3 is configured to determine a control variable vector comprising control variables for controlling the power device 5. For this purpose, it has a first optimization module 7, which is configured to determine a first optimized control variable vector by means of a first, non-evolutionary algorithm, and a second optimization module 9, which is configuredto receive the first optimized manipulated variable vector from the first optimization module 7 and to determine a second optimized manipulated variable vector by means of a second, evolutionary algorithm using the first optimized manipulated variable vector, as well as a control module 11 configured to determine the control manipulated variable vector as a function of the second optimized manipulated variable vector. In particular, the control device 3 is configured to carry out an embodiment of a method for operating the power device 5, described below. Fig. 2 shows a schematic representation of an embodiment of a method for operating the power device 5 in the form of a flowchart. Identical and functionally identical elements are provided with the same reference numerals in all figures.so that in this respect, reference is made to the preceding description. Within the scope of the method, in particular, a first optimized manipulated variable vector is determined using a first, non-evolutionary algorithm, a second optimized manipulated variable vector is determined using a second, evolutionary algorithm using the first optimized manipulated variable vector, wherein a control manipulated variable vector is determined as a function of the second optimized manipulated variable vector, and wherein the power device is operated with the control manipulated variable vector. In one embodiment, the first optimized manipulated variable vector is determined using a gradient method as the first, non-evolutionary algorithm. In particular, in a first step S1—by the first optimization module 7—the first optimized manipulated variable vector 1SV is determined using a first, non-evolutionary algorithm.in particular calculated. For this purpose, the first optimization module 7 uses in particular operating parameters BP of the power device 5, wherein the operating parameters BP are included in the determination of the first optimized manipulated variable vector 1SV. The operating parameters BP can in particular be received by the first optimization module 7 or another module of the control device 3 - in particular via at least one suitable interface, for example from sensors - or they can be determined, in particular calculated or simulated, by the first optimization module 7 or another module of the control device 3 itself, in particular based on measured values ​​determined by at least one sensor. The operating parameters BP are preferably selected from a group consisting of an ambient temperature, an ambient air pressure, a target value for the operation of the power device 5, a limit value for the operation of the power device 5,at least one environmental boundary condition, and a combination of at least two of the mentioned parameters. The operating parameters BP can also include further parameters not explicitly mentioned here. Preferably, the first optimized manipulated variable vector 1SV is determined by the first optimization module 7 for a predetermined prediction interval. In particular, the first optimized manipulated variable vector 1SV is determined for the end of the predetermined prediction interval. In particular, the first optimized manipulated variable vector 1SV is calculated using a gradient method. Alternatively or additionally, the first optimized manipulated variable vector 1SV is calculated in a stationary manner. Preferably, the first optimized manipulated variable vector 1SV is calculated using a model-based, predictive method, in particular using the same method,which is also used by the control module 11 to calculate the control manipulated variable vector; however, the first optimized manipulated variable vector 1SV is preferably calculated in a stationary manner, while the dynamics of the power device 5 are included in the calculation of the control manipulated variable vector. In a second step S2, the first optimized manipulated variable vector 1SV is received by the second optimization module 9 from the first optimization module 7, and a second optimized manipulated variable vector 2SV is determined, in particular calculated, by the second optimization module 9 using a second, evolutionary algorithm using the first optimized manipulated variable vector 1SV. Preferably, a predetermined number of test manipulated variable vectors are generated for this purpose, and the first optimized manipulated variable vector 1SV is used as a further test manipulated variable vector.so that a first set of manipulated variable vectors with the cardinality of a predetermined total number is formed. The first set of manipulated variable vectors comprises the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector 1SV as a further test manipulated variable vector. The second optimized manipulated variable vector 2SV is determined from the first set of manipulated variable vectors using the second, evolutionary algorithm. The operating parameters BP are again used to determine the second optimized manipulated variable vector 2SV, in particular for calculating a cost function.by means of which the test manipulated variable vectors are evaluated. The determination of the second optimized manipulated variable vector 2SV is explained in more detail in connection with Figure 3. In particular, as shown by a dashed arrow, the second optimized manipulated variable vector 2SV determined in the second step S2 is transmitted to the first optimization module 7. The first step S1 is repeated, in particular, with a first timing cycle having a first sampling time, wherein a new first optimized manipulated variable vector 1SV is determined in each case based on the currently transmitted second optimized manipulated variable vector 2SV. As also shown by a dashed arrow, each newly determined first optimized manipulated variable vector 1SV is preferably transferred to the control module 11 in order to determine the control manipulated variable vector AV. The second step S2 is preferably repeated with a second timing cycle having a second sampling time,wherein the first sampling time is in particular shorter than the second sampling time. Preferably, if no new second optimized manipulated variable vector 2SV is currently available, the respective current, new first optimized manipulated variable vector 1SV is used by the control module 11 instead of the last received second optimized manipulated variable vector 2SV to determine the control manipulated variable vector AV. Starting from the second step S2, a third step S3 checks whether a predetermined termination condition is met. If this is not the case, the second step S2 is repeated, wherein a new second optimized manipulated variable vector 2SV is determined using a new first optimized manipulated variable vector 1SV, which is again currently received from the first optimization module 7. The second and third steps S2 and S3 are repeated until the predetermined termination condition is met in the third step S3. If this is the case,The method ends in a fourth step S4. The repetition of the second and third steps S2 and S3 preferably takes place with the second time cycle having the second sampling time. In particular, the first sampling time is shorter than the second sampling time. Preferably, the first sampling time is 10 ms, and the second sampling time is 100 ms. The first step is therefore carried out more frequently than the iteration of the second and third steps S2 and S3. Simultaneously with the iteration of the second and third steps S2 and S3 described in the previous paragraph, each newly determined second optimized manipulated variable vector 2SV in the current second step S2 is always transferred to the control module 11, which in a fifth step S5 determines a control manipulated variable vector AV depending on the second optimized manipulated variable vector 2SV. If no new second optimized manipulated variable vector 2SV is currently received,However, a new first optimized manipulated variable vector 1SV, the control manipulated variable vector AV is determined as a function of the new first optimized manipulated variable vector 1SV. Preferably, the control manipulated variable vector AV is determined by the control module 11 using a model-based, predictive method, penalizing a deviation of the control manipulated variable vector AV from the second optimized manipulated variable vector 2SV - or alternatively, the new first optimized manipulated variable vector 1SV. In particular, a manipulated variable vector trajectory is preferably determined for the predetermined prediction interval, and a current manipulated variable vector of the manipulated variable vector trajectory is determined as the control manipulated variable vector AV. In particular, the control manipulated variable vector AV, in particular the manipulated variable vector trajectory,preferably calculated taking into account the dynamics of the power device 5. In a sixth step S6, the power device 5 is then operated using the control variable vector AV determined in the fifth step S5, in particular by controlling the power devices 5 with the manipulated variables included as components of the control variable vector AV. Figure 3 shows a schematic representation of a detail of the method according to Figure 2. In particular, the determination of the second optimized manipulated variable vector 2SV in the second step S2 is explained in more detail with reference to Figure 3. First, in a first second step S2.1, the first optimized manipulated variable vector 1SV and the operating parameters BP are received. In a second second step S2.2, the predetermined number – G-1 – of test manipulated variable vectors TSV is randomly generated.and the first set of manipulated variable vectors with the cardinality of the predetermined total number G is formed from the predetermined number G-1 of randomly generated test manipulated variable vectors TSV and the first optimized manipulated variable vector 1SV received from the first optimization module 7 as a further test manipulated variable vector TSV. In a third second step S2.3, each test manipulated variable vector TSV of the first set is evaluated using a cost function—in particular, taking into account the operating parameters BP. In a fourth second step S2.4, a second set of descendant manipulated variable vectors NSV with the cardinality of the predetermined total number G is formed from a predetermined parent number E of the test manipulated variable vectors TSV of the first set that were best evaluated using the cost function. In a fifth second step S2.5, the descendant manipulated variable vectors NSV are mutated, with the exception of a predetermined subset U of elite manipulated variable vectors.wherein a third set of manipulated variable vectors with the cardinality of the predetermined total number G is formed, wherein the third set comprises the mutated offspring manipulated variable vectors mNSV and the elite manipulated variable vectors. Finally, in a sixth second step S2.6, the test manipulated variable vector TSV best evaluated using the cost function is determined as the second optimized manipulated variable vector 2SV. Returning to Figure 2, during the iteration of the second and third steps S2 and S3, a new first optimized manipulated variable vector 1SV is always received from the first optimization module 7 in each further execution of the first second step S2.1 by the second optimization module 9, and a fourth set of manipulated variable vectors with the cardinality of the predetermined total number G is formed in each further execution of the second second step S2.2.by replacing a manipulated variable vector of the third set with the new first optimized manipulated variable vector 1SV. In particular, to form the fourth set, the lowest-ranking, worst-rated manipulated variable vector is replaced by the new first optimized manipulated variable vector 1SV. The newly formed fourth set is then used as the new first set. The further iteration steps, in particular steps S2.3 to S2.6, are then repeated with the fourth set formed in the second step S2.2 as the new first set, in particular until the predetermined termination condition is met in the third step S3.

Claims

CLAIMS 1. Control device (3) for a power device (5), wherein the control device (3) is configured to determine a control manipulated variable vector comprising manipulated variables for controlling the power device (5), with ˗ a first optimization module (7) configured to determine a first optimized manipulated variable vector by means of a first, non-evolutionary algorithm, ˗ a second optimization module (9) configured to receive the first optimized manipulated variable vector from the first optimization module (7) and to determine a second optimized manipulated variable vector by means of a second, evolutionary algorithm using the first optimized manipulated variable vector, and ˗ a control module (11) configured to determine the control manipulated variable vector as a function of the second optimized manipulated variable vector.The control device (3) according to claim 1, wherein the first optimization module (7) is configured to determine the first optimized manipulated variable vector using a gradient method as the first, non-evolutionary algorithm.

3. The control device (3) according to any one of the preceding claims, wherein the second optimization module (9) is configured to generate a predetermined number of test manipulated variable vectors, to use the first optimized manipulated variable vector as a further test manipulated variable vector, so that a first set of manipulated variable vectors with the cardinality of a predetermined total number is formed, wherein the first set of manipulated variable vectors comprises the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector, and to determine the second optimized manipulated variable vector from the first set of manipulated variable vectors using the second, evolutionary algorithm.Control device (3) according to one of the preceding claims, wherein the second optimization module (9) is configured to a) generate a first set having the power of a predetermined total number of manipulated variable vectors from a predetermined number of randomly generated test manipulated variable vectors and that received from the first optimization module (7). to form a first optimized manipulated variable vector as a further test manipulated variable vector, wherein the first set of manipulated variable vectors comprises the predetermined number of test manipulated variable vectors and the first optimized manipulated variable vector as a further test manipulated variable vector; b) to evaluate each test manipulated variable vector of the first set using a cost function; c) to form a second set of descendant manipulated variable vectors with the cardinality of the predetermined total number from a predetermined parent number of the test manipulated variable vectors of the first set that were best evaluated using the cost function; d) to change the descendant manipulated variable vectors with the exception of a predetermined minority of elite manipulated variable vectors, wherein a third set of manipulated variable vectors with the cardinality of the predetermined total number is formed, wherein the third set comprises the changed descendant manipulated variable vectors and the elite manipulated variable vectors,and e) to determine the test manipulated variable vector best evaluated by means of the cost function as the second optimized manipulated variable vector.

5. The control device (3) according to claim 4, wherein the second optimization module (9) is further configured to f) receive a new first optimized manipulated variable vector from the first optimization module (7) and to form a fourth set of manipulated variable vectors with the cardinality of the predetermined total number by replacing a manipulated variable vector of the third set with the new first optimized manipulated variable vector, and g) to repeat steps b) to f) with the fourth set obtained in step f) as the first set, in particular until a predetermined termination condition is met.

6. The control device (3) according to one of the preceding claims, wherein the first optimization module (7) is configured toin particular for the end of the predetermined prediction interval.

7. The control device (3) according to one of the preceding claims, wherein the control module (11) is configured to determine the control manipulated variable vector using a model-based, predictive method while penalizing a deviation of the control manipulated variable vector from the second optimized manipulated variable vector.

8. The control device (3) according to one of the preceding claims, wherein the control module (11) is configured to determine a manipulated variable vector trajectory for the predetermined prediction interval and to determine a current manipulated variable vector of the manipulated variable vector trajectory as the control manipulated variable vector.

9. The control device (3) according to one of the preceding claims, wherein the first optimization module (7) is configured to calculate the first optimized manipulated variable vector in a steady-state manner.Control device (3) according to one of the preceding claims, wherein the control module (11) is configured to calculate the control manipulated variable vector, in particular the manipulated variable vector trajectory, taking into account dynamics of the power device (5).

11. Power arrangement (1) comprising a power device (5) and a control device (3) according to one of claims 1 to 10.

12. Method for operating a power device (5), wherein ˗ a first optimized manipulated variable vector is determined by means of a first, non-evolutionary algorithm, ˗ a second optimized manipulated variable vector is determined by means of a second, evolutionary algorithm using the first optimized manipulated variable vector, wherein ˗ a control manipulated variable vector is determined as a function of the second optimized manipulated variable vector, and wherein ˗ the power device (5) is operated with the control manipulated variable vector.

Images